LUMORA guide

Workflows

Users think in tasks. Not features.

Find receipts

1. Enable OCR
2. Index library
3. Search: receipt

Find vacation photos

1. Search: beach sunset
2. Or try family vacation / mountain hike

Remove duplicates

1. Open Duplicates
2. Review matches
3. Keep best version

Rediscover old memories

1. Open Memories
2. Browse events
3. Explore Timeline

Search cookbook

Example natural-language queries (CLIP installed):

• dog running
• birthday party
• beach sunset
• snow mountain
• pizza
• family picnic
• wedding

Install & run

Prefer a ready-made build? Use the Install & usage page or Download for platform installers.

To build from source you need Bun, a stable Rust toolchain, and (on macOS) Xcode Command Line Tools.

bun install
bun run tauri dev      # desktop app + Vite
bun run tauri build    # release bundle

• Ensure ~/.cargo/bin is on your PATH.
• For video thumbnails, install ffmpeg / ffprobe so they are on PATH. Without them, videos still import; thumbs fall back to a placeholder.
• Contributor notes live in CONTRIBUTING.md; architecture decisions in docs/adr/.

Import your library

1. Open Home → Import photos, or drop a folder onto the window.
2. Originals stay on disk. LUMORA stores metadata, thumbs, and optional AI data in app data.
3. Add Watched folders so new files appear automatically (Settings → Library, or the Watched folders view).
4. Import timing is logged locally (Activity + Developer → import analytics). Nothing is uploaded.

Browse & organise

• Library — virtualised grid of everything.
• Timeline — year / month browsing.
• Albums — manual collections; drag photos in.
• Tags, ratings, colour labels, favourites — from the selection bar or viewer.
• Smart collections — Videos, RAW, Screenshots (path contains “screenshot”), Selfies, Panoramas, Documents, Receipts.
• Open the info panel on any item for EXIF, tags, auto-tags, captions, faces, and OCR text.

Search

The toolbar search mixes full-text and (when models are installed) semantic results.

• Plain text matches filename, tags, camera, OCR, people, auto-tags, captions.
• Filters such as camera:iphone, rating>3, before:2024-01-01.
• Natural language (“dog on a beach”) uses on-device CLIP when the semantic model is installed and enabled.
• Recent queries appear as toolbar hints and under Discover.

On-device AI

All models are optional. Install them from Settings → AI. Downloads are user-initiated and checksum-verified.

Semantic search

CLIP ViT-B/32 — find by description.

OCR

PaddleOCR PP-OCRv5 (default) or PP-OCRv6 small — text in screenshots and documents. Older RapidOCR PP-OCRv4 / v3 remain available in the model library.

Faces / People

InsightFace (buffalo_l / buffalo_s). Only faces that are ≥80% in-frame and ≥80% front-facing are kept. Name, merge, ignore in People.

Auto-tags

MobileNetV4 ImageNet labels. Shown in the info panel and searchable. Enable Object detection / auto-tags.

Image captions

Florence-2 Base — short on-device descriptions. Shown in the info panel and searchable. Enable Image captions (~277 MB download).

Model library

Switch backends per capability, install the active option, and re-run processing when you change models. For Auto-tags and Semantic search you can also Import local model… — pick ONNX files from disk; LUMORA evaluates compatibility and only installs if the graph matches. (Browsing Hugging Face from the app is not available yet.)

You can clear embeddings, OCR, faces, auto-tags, or captions anytime without deleting original files.

Duplicates & blurry cleanup

Open Duplicates in the sidebar. Sections are collapsible so you can scan quickly.

• Exact — identical SHA-256. Use Clean up all exact — keep 1 per group, or keep a specific file per group.
• Near — perceptual aHash with Hamming ≤ 2. Review each group; there is no bulk delete for near matches (false positives can still happen with a coarse hash).
• Blurry — Laplacian variance ≤ 80. Trash one or trash all listed after review. Soft-focus portraits can score low too.

Everything goes to Trash first — restore or undo with ⌘Z / Ctrl+Z.

Places & people

• Places groups photos by GPS EXIF using offline reverse geocoding (bundled GeoNames) and shows them on an offline geometry map. No map tile network calls.
• People lists face clusters after face models are installed. Name people to make them searchable.

Encrypted vault

Move private items into the Locked folder. Keys use Argon2id; content uses XChaCha20-Poly1305. Store your recovery code safely — it is the only unlock path if you forget the password.

Privacy defaults

• No telemetry, analytics, or cloud photo APIs.
• Network is used when you download models you chose, and optionally when checking for app updates (Settings → Updates).
• Diagnostic logs stay on disk (Developer page).
• Derived AI data is rebuildable and deletable; originals stay put.

FAQ

Keyboard shortcuts